AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

AMDP: Asynchronous Multi-Directional Pipeline Parallelism for Large-Scale Models Training

Ling Chen, Houming Wu, Wenjie Yu

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add zero redundancy optimizer (ablation)AMDP · GPT-style, 8 GPUs, d=8, b=128throughput (ktokens/s): 80.3 -> 83.7Improved
add zero redundancy optimizer (ablation)AMDP · BERT-style, 8 GPUs, d=8, b=128throughput (ktokens/s): 82.7 -> 86.1Improved
add zero redundancy optimizer (ablation)AMDP · GPT-style, d=4 and d=8average GPU memory: without ZeRO memory rises by 11.4% and 19.4%Improved
add zero redundancy optimizer (ablation)AMDP · BERT-style, d=4 and d=8average GPU memory: without ZeRO memory rises by 9.7% and 15.4%Improved
replace PipeDream-2BW with AMDP multi-directional schedulingpipeline parallelism · GPT-style, 8 GPUs, d=8, b=128throughput (ktokens/s): 71.6 -> 83.7Improved
replace PipeDream-2BW with AMDPpipeline parallelism · BERT-style, 8 GPUs, d=8, b=128throughput (ktokens/s): 75.8 -> 86.1Improved
replace PipeDream-2BW with AMDP on 16 GPUspipeline parallelism · GPT-style, 16 GPUs, d=16throughput (ktokens/s): 110.1 -> 124.3Improved
replace vNAG with AMDPpipeline parallelism · Model-1 vs vNAGthroughput (ktokens/s): 47.1 -> 58.5Improved
replace vNAG with AMDPpipeline parallelism · Model-2 vs vNAGthroughput (ktokens/s): 11.1 -> 18.2Improved
replace vNAG with AMDPpipeline parallelism · Model-1 vs vNAGtraining loss at 50k iterations: 5.12 -> 4.09Improved
replace asynchronous PipeDream-2BW with AMDP (bounded mismatch)pipeline parallelism · GPT-style validation at 40k iterationsperplexity: 21.7 -> 2.06Improved
replace asynchronous PipeDream-2BW with AMDP (bounded mismatch)pipeline parallelism · BERT-style validation at 40k iterationsperplexity: 65.7 -> 1.7Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy